A hypergraph-based hybrid graph convolutional network for intracity human activity intensity prediction and geographic relationship interpretation

Human activity intensity prediction, i.e., estimating the dynamic population distribution, is crucial to many location-based applications, particularly intelligent transportation systems and urban planning. Geographical artificial intelligence (GeoAI) is an emerging field that has contributed many new modeling approaches for spatiotemporal phenomena, especially spatiotemporal graph neural networks (ST-GNNs). However, there are two challenges to the prediction of human activity intensity. On the one hand, accurate predictions require modeling not only pairwise spatial relationships, e.g., distance and spatial flow but also complex high-order relationships, such as clusters and groups. On the other hand, it is not appropriate to incorporate prior knowledge on the spatial structure of data with subjective assumptions using graphs with predefined topologies or weights. To mitigate such challenges, a novel deep learning-based model was proposed, namely, the Hypergraph-based Hybrid Graph Convolutional Network (HyGCN). In our study, high-order relationships are modeled by hypergraph convolutions, and pairwise relationships are captured in graph convolutions. A spatial fusion layer is added to fuse the pairwise and high-order relationships. Then, adaptive graph and hypergraph generation methods are proposed to interpret edge weights and understand the learned geographic relationships in accordance with principles. After systematic experiments on large-scale mobility data, our results demonstrate that HyGCN outperforms state-of-the-art models. The learned geospatial knowledge can be disentangled and explained from the perspectives of geographic principles. This study offers new insights for not only revisiting geographic laws in GeoAI but also utilizing spatiotemporal knowledge in deep learning models.